Chollet critiques the overreliance on scaling pre-trained models, arguing instead for systems that can adapt to novel situations through abstraction and reasoning. He presents the ARC benchmark as a more accurate measure of general intelligence, showing how even advanced models still fall short of human-level performance. He also introduces upcoming versions of ARC designed to test interactivity and agency. Finally, he outlines a new approach to AI development that combines deep learning with algorithmic components and program search, aiming to build systems capable of invention and real-world problem-solving.